Academic Radar

Using this framework, we reformalize existing attacks, such as indirect prompt injection, direct prompt injection, jailbreak, task drift, and memory poisoning,...

To this end, we propose four security properties that capture contextual security for LLM agents: *task alignment* (pursuing authorized objectives), *action alignment* (individual actions serving those objectives), *source authorization* (executing commands from authenticated sou...

Five-layer defense-in-depth architecture aligned with the agent lifecycle. ... Li (2024) Jailbreak attacks and defenses against large language models: a survey....

The System Prompt Is the Attack Surface: How LLM Agent Configuration Shapes Security and Creates Exploitable Vulnerabilities ... LLM alignment. Instruction...

Despite these advances, LLMs remain highly vulnerable to jailbreak attacks Shen et al. (2024); Wang et al. (2025a) , where adversaries craft malicious queries...

Abstract. Deploying large language models (LLMs) as autonomous browser agents exposes a significant attack surface in the form of Indirect Prompt Injection...

The ASR alignment score is calculated by aggregating the outcomes of multiple queries in the target dataset, then we update the template using the previous ones...

At a high level, our approach consists of (i) observing benign and adversarial interactions with an agent to derive a set of representative execution traces, (ii) abstracting these traces into functionality graphs that capture recurring execution flows and data provenance, and (i...

Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning Large Language Models (LLMs) with human values. However, RLHF has been continuously...

Its key idea is online but off-policy RLHF, updating models with samples from previous iterations. While effective, it remains a typical off-policy strategy and...

Recent advances in large language models (LLMs) have been driven by reinforcement-learning-based post-training, which requires multiple rollouts with...

In this work, we present a unified framework that enables LLM agents to learn more effectively through broader exploration by jointly updating their parametric...

Standard reinforcement learning (RL) for large language model (LLM) agents typically optimizes extrinsic rewards, prioritizing isolated task completion over...

Supervised Fine-Tuning versus Reinforcement Learning: A Study of Post-Training Methods for Large Language Models Haitao Jiang1, Wenbo Zhang2, Jiarui Yao3, Hengrui Cai2, Sheng Wang4, Rui Song5 1 North Carolina State University 2 UC Irvine 3 University of Illinois Urbana-Champaign ...

CausalRM: Causal-Theoretic Reward Modeling for RLHF from Observational User Feedbacks Hao Wang1,2, Licheng Pan2,3, Zhichao Chen1, Chunyuan Zheng1, Zhixuan Chu3, Xiaoxi Li2, Yuan Lu2, Xinggao Liu3, Haoxuan Li1∗, Zhouchen Lin1∗ 1Peking University, 2Xiaohongshu Inc., 3Zhejiang Unive...

Fig.[4](https://arxiv.org/html/2603.12671v1#S6.F4 "Figure 4 ‣ VI-A6 Performance Metrics ‣ VI-A Parameter Settings ‣ VI Evaluation ‣ HyGra: Accelerating Network-State Simulation for LLM Training in DCNs via Adaptive Packet-Flow Granularity") compares the simulation runtime of HyGr...

Centrifuge features two main design points: 1) At the algorithm level, Centrifuge filters the activations in the memory-efficient attention backward kernel to amplify sparsity (§4.1); 2) At the system level, Centrifuge transforms the sparse GEMM to dimension-reduced dense GEMM th...

Mixture-of-Experts (MoE) architecture offers enhanced efficiency for Large Language Models (LLMs) with modularized computation, yet its inherent sparsity poses significant hardware deployment challenges, including memory locality issues, communication overhead, and inefficient co...

March 17, 2026 Towards Next-Generation LLM Training: From the Data-Centric Perspective Hao Liang1, Zhengyang Zhao1, Zhaoyang Han1, Meiyi Qiang1, Xiaochen Ma5, Bohan Zeng1, Qifeng Cai1, Zhiyu Li2,4, Linpeng Tang2,3, Weinan E1,2,3, Wentao Zhang‡,1,2,3 1Peking University, Beijing, C...

Therefore, the overall training throughput of LLM distillation is often bounded by the inefficient execution of the teacher model, especially when the teacher...

Communication links are established based on spatial proximity, forming a connected topology as shown in Fig. 6 ‣ Appendix G Simulation Environments and Task Configurations ‣ MACRO-LLM: LLM-Empowered Multi-Agent Collaborative Reasoning under Spatiotemporal Partial Observability")...

Abstract. Lossless model compression holds tremendous promise for alleviating the memory and bandwidth bottlenecks in bit-exact Large Language Model (LLM)...

Takeaway: Distributed training exhibits non-linear scal- ability and diminishing returns from the interplay between computation, communication, and memory...

To facilitate audio generation via the LLM, we developed a specialized speech tokenizer based on the WavLM-large (chen2022wavlm) , augmented with a single...

Beyond audio-to-text models, an emerging research direction explores multi-modal LLMs designed for interactive scenarios that require both audio understanding...

We further design a arXiv:2603.16280v1 [cs.SD] 17 Mar 2026 Flow-Matching Transformer Speaker Encoder Flan-T5 Projector Text Prompt ConvNext V2 Blocks Target Transcription c Vocoder Timbre Embed G Cross-attention Noise Duration Predictor Trainable Frozen Infer Chosen Train & Infer...

Prior work reports strong emotion recognition accuracy on human speech using both discriminative SER models and generative Speech Large Language Models (SLMs) (Koh and Dubnov, 2021; Sadok et al., 2023; Ma et al., 2024; Wu et al., 2025a), yet generating speech that reliably expres...

AT-ADD: The Grand Challenge on All-Type Audio Deepfake Detection Yuankun Xie* xieyuankun@cuc.edu.cn Communication University of China Beijing, China Haonan Cheng* haonancheng@cuc.edu.cn Communication University of China Beijing, China Xiaoxuan Guo Communication University of Chin...

On the Fish Audio Instruction Benchmark, Fish Audio S2 achieves an overall tag-activation rate of 93.3% and an overall quality score of 4.51/5.0 across English and Chinese, as evaluated by Gemini 3 Pro. To accelerate research and lower the barrier to high-quality TTS development,...

We introduce DashengTokenizer, a unified continuous audio tokenizer designed for both understanding and generation across speech, music, and environmental sound...

In this paper, we evaluate the adaptability of CFG strategies originally developed for image generation to speech synthesis and extend separated-condition CFG...